A Performance Study of Event Processing Systems

Event processing engines are used in diverse mission-critical scenarios such as fraud detection, traffic monitoring, or intensive care units. However, these scenarios have very different operational requirements in terms of, e.g., types of events, queries/patterns complexity, throughput, latency and number of sources and sinks. What are the performance bottlenecks? Will performance degrade gracefully with increasing loads? In this paper we make a first attempt to answer these questions by running several micro-benchmarks on three different engines, while we vary query parameters like window size, window expiration type, predicate selectivity, and data values. We also perform some experiments to assess engines scalability with respect to number of queries and propose ways for evaluating their ability in adapting to changes in load conditions. Lastly, we show that similar queries have widely different performances on the same or different engines and that no engine dominates the other two in all scenarios.